Research On Detection Method Of Steel Strip Surface Defects Based On Deep Learning

Strip steel is a steel product commonly used in industrial production.During the production process,due to the influence of rolling equipment and production technology,there will be some defects on the surface of the strip steel,which will affect the use of subsequent strip steel.Therefore,it is a key issue to improve the production efficiency and production quality of strip steel to detect defects quickly and accurately.Traditional inspection methods usually include manual inspection,non-destructive inspection and machine vision inspection.But these detection methods are usually inefficient,costly and poor in generalization performance.The detection method based on deep learning can automatically extract features,and the detection accuracy is also higher.Therefore,the main research topic of this thesis is to use computer vision technology based on deep learning to solve the problem of surface defect detection of strip steel.Compared with the images in natural scenes,the surface defects of strip steel have the characteristics of small differences between classes and large differences within classes.According to this characteristic and the actual needs of strip steel production,this thesis designs two defect detection methods,namely,the defect detection method based on Transformer and the defect detection method based on context fusion and feature refinement.The main work contents and innovations are as follows:(1)Transformer-based defect detection method.In this method,an end-to-end defect detection method is designed according to the characteristics of the strip surface defect.It mainly uses Anchor-free technology based on YOLOX,and uses MobileViT based on CNN and Transformer as a feature extraction network.On this basis,an Improved Path Aggregation Network(IPAN)is designed,which can optimize the fuzzy problem of positive sample classification in anchor-free technology while introducing contextual information to improve the detection accuracy and recall rate of the model.And the introduction of CIoU Loss in the loss function improves the positioning accuracy of the model.Finally,when the input size is 640×640,this method achieves 70.52%mAP on the GC10-DET dataset and 76%mAP on the NEU-DET dataset at a detection speed of 14FPS.(2)Defect detection method based on context fusion and feature refinement.This method is proposed in view of the fact that the accuracy and real-time performance of the current strip surface defect detection are difficult to meet the actual production needs.In the method,the SSD with good detection accuracy and speed is used as the overall framework of the target detection.At the same time,according to the feature of strip surface defect,a Context Fusion Structure(CFS)and a Feature Refinement Module(FRM)are designed.The context fusion structure can improve the ability of the model to capture the surface defects of the strip while ensuring the inference speed.The feature refinement module uses the attention mechanism to guide the model to perform context fusion,filter out semantic conflicts and redundancy caused by multiscale information fusion,and improve the accuracy of target detection,thereby further improving the accuracy of SSD detection of steel surface defects.Finally,when the input size is 300×300,79.5%mAP is obtained at a detection speed of 71FPS on the NEU-DET data set,which can meet the real-time performance in actual production and achieve the highest accuracy.